An USPCD multi-feature fusion method for bridge underwater structure identification

Recent achievements in 3D sonar enable centimeter-level resolution of underwater bridge structure measurement. However, this theoretical accuracy cannot be directly obtained from the raw underwater sonar point cloud data (USPCD) due to the significant scattering characteristics, non-uniform density distribution, and a large amount of noise that seriously pollute and mask the structural features of interest. Therefore, we propose an advanced USPCD multi-feature extraction method for bridge pile morphological inspection, providing valuable substructure in-service condition. This method is based on two-dimensional space, fusing point cloud density and edge features to identify the edge points of piles, and using spectral clustering methods to classify and locate pile foundations. Subsequently, coarse segmentation based on minimum cuts and fine extraction based on statistical methods are implemented in sequence to achieve complete segmentation of pile foundations and effective filtering of various environmental noises, thereby realizing high-performance extraction of piles with different distribution patterns and different degrees of incompleteness. The efficacy of the method is validated through an application to deep-water group pile foundations from two bridges across the Yangtze River. The results demonstrate the strong generalizability of the proposed method, with instance segmentation accuracy metrics exceeding 0.88 across all validation cases. Specifically, for relatively complete pile groups, the P ¯ , R ¯ , F ¯ 1 , and mIoU metrics all surpass 0.89. Even for pile groups with an incompleteness rate of 90%, the minimum P ¯ still reachPes 0.863, thereby satisfying the requirements for detailed inspection of underwater bridge substructures.